About NVIDIA builder stack

NVIDIA gives your team an open-model layer for the hackathon. Two ways in, same simple setup:

• **build.nvidia.com:** NVIDIA's hosted API Catalog: 100+ models (reasoning, vision, retrieval/embeddings, speech, guardrails) served as NIM microservices on NVIDIA infrastructure. Free for NVIDIA Developer Program members, one nvapi- key, one REST endpoint for every model.
• **Nemotron on OpenRouter:** NVIDIA's Nemotron family of open models (open weights, data, and recipes), available free through OpenRouter's unified API. Built for agentic AI with reasoning—the kind of multi-step, tool-using workflows legaltech teams tend to prototype.

Both are simple REST APIs, and the Friday workshop walks you through onboarding so you arrive Saturday ready to build, not configure.

<aside> <img src="notion://custom_emoji/9067698e-8c48-4c31-9814-0f22ae0539d7/378c01c2-41b2-802a-941e-007a5d32fd71" alt="notion://custom_emoji/9067698e-8c48-4c31-9814-0f22ae0539d7/378c01c2-41b2-802a-941e-007a5d32fd71" width="40px" />

Pre-hackathon workshop: Building with NVIDIA

</aside>

👀 Why it's useful at LLMxLaw Hackathon

Some ideas on where to use the NVIDIA stack:

• Open-model inference layer for agents. Run your agentic legal workflows on Nemotron's reasoning models—free on OpenRouter, $0/token. Reach for this when you want an open, customisable model rather than a closed API, and don't want token cost to cap how much you experiment.

  Useful in combination with: Neo4j, EU Publications Office data API

• Long-context document reasoning. Nemotron 3 Super carries a very large (1M-token) context window, so large case files, contract sets, or bundles can be reasoned over with less aggressive chunking. Useful in combination with: Neo4j (graph retrieval), Perplexity (live search)

• Multimodal intake. Nemotron 3 Nano Omni reads text, images, video, and audio in a single loop—useful for scanned filings, exhibits, or recorded proceedings rather than text alone. Useful in combination with: Lovable / Momen (frontend)

• Build a full stack from one catalog. build.nvidia.com hosts embeddings/retrieval, guardrails, and speech models alongside LLMs, so you can assemble a RAG or agent pipeline from a single endpoint instead of stitching providers together. Useful in combination with: GCP

🔑 How access works

There's nothing to install—access is two free sign-ups:

1. build.nvidia.com: join the free NVIDIA Developer Program (work or personal email, no credit card), then generate an API key (prefixed nvapi-). That key opens the full model catalogue via a single REST endpoint.
2. OpenRouter: create a free OpenRouter account to call the Nemotron :free endpoints. Same setup; set the model string (e.g. nvidia/nemotron-3-super-120b-a12b:free) and your OpenRouter key.

👉 Get started!

Both are plain REST APIs, so any HTTP client or SDK works. A minimal call to a free Nemotron model on OpenRouter:

curl <https://openrouter.ai/api/v1/chat/completions> \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/nemotron-3-super-120b-a12b:free",
    "messages": [{"role": "user", "content": "Hello"}]
  }'

For build.nvidia.com, open any model on the catalog and copy the ready-made code snippet shown on its page; it includes the exact endpoint and model string.